in visual-form recognitionlevels.psych.umn.edu/pubs_index/pubs_web_pdfs/m_b_97.pdf ·...

In S. Christman (Ed.) (1997), Cerebral Asymmetries in Sensory and 1Perceptual Processing. Amsterdam: Elsevier.

COMPUTATIONAL ANALYSES AND HEMISPHERIC ASYMMETRIES

IN VISUAL-FORM RECOGNITION

Chad J. Marsolek E. Darcy Burgund

University of Minnesota University of Minnesota

Running head: VISUAL-FORM RECOGNITION

Please send correspondence to: C. J. Marsolek Phone: 612-624-1597

Department of Psychology Fax: 612-626-2079

75 East River Road

University of Minnesota E-mail: marso002@

Minneapolis, MN 55455 gold.tc.umn.edu

Visual-Form Recognition 2

COMPUTATIONAL ANALYSES AND HEMISPHERIC ASYMMETRIESIN VISUAL-FORM RECOGNITION

Much of contemporary theory in cognitive neuroscience adheres to a theoretical framework inwhich behavioral abilities are understood as arising from the operations of and interactions betweenrelatively independent processing subsystems of the brain. Although these subsystems may be only weaklymodular (e.g., Farah, 1994), a primary goal for the field is to discover the broad architecture ofinterconnected subsystems and to pinpoint the particular functions performed by the various subsystems.In other words, the goal is to understand how the brain is “carved at its joints” (e.g., Kosslyn & Koenig,1992) or, perhaps more appropriately given weak modularity, “stretched at its interconnections” intorelatively independent functional entities. This quest for component subsystems is especially importantin understanding functional hemispheric asymmetries. Theories that are cast in terms of how componentsubsystems operate more or less effectively in different hemispheres are providing more powerful andcompelling explanations than those that rely on general principles or fundamental dichotomies (seeHellige, 1993a).

In order to theorize about the function of any system, one must characterize how inputs aremapped to outputs by the system. Inputs, outputs, and the couplings between the two are needed todescribe any function; if one of these three elements is missing or vague, no matter which one, a functionper se cannot be described. Hence, if one’s goal is to delineate and understand the function of a neuralprocessing subsystem, one must theorize not only about what it produces as output, but also about whatit accepts as input and which inputs will be used to signal the production of which outputs. Thissomewhat obvious point helps to illuminate some of the important questions to consider when doing thedifficult job of stretching the brain at its interconnections, functionally speaking.

An effective strategy to follow when hypothesizing about neural processing subsystems is toconsider clues about inputs, outputs, and their mappings from the perspectives of different levels ofexplanation (Marr, 1982). Useful computational clues come from considerations of the goals that shouldbe satisfied by the relevant subsystems, what information is available to the relevant subsystems to helpachieve these goals, and what sort of strategy would be useful for achieving the appropriate goals giventhe available information. Useful implementational clues include aspects of the underlying physicalsubstrate that suggest how inputs may be represented, where the inputs come from, how outputs may berepresented, where the outputs are sent to, and how the mechanism may operate to map inputs tooutputs.

In this chapter, we consider such clues to hypothesize about the subsystems involved in visual-form recognition. First, we theorize that two relatively independent subsystems underlie different aspectsof visual-form recognition and that each subsystem operates more effectively in one cerebral hemispherethan in the other. Then, we summarize the results from behavioral studies that support the separatesubsystems theory. Next, we offer computational analyses and describe results from a computationalmodeling study that illuminate the contradictory natures of the internal processing strategies that thesesubsystems may use. Finally, we summarize the results from additional behavioral studies that supportthe computational theory.

VISUAL-FORM SUBSYSTEMSVisual-form recognition is an essential human ability. By most accounts, it entails the activation

of a previously stored visual-form representation that corresponds best to the currently processed inputform. Generally, the neural mechanisms involved in this ability appear to operate in occipital-temporaland inferior-temporal cortex of the brain (e.g., Buckner et al., 1995; Petersen & Fiez, 1993; Schacter et al.,


1995; Sergent, Ohta, & MacDonald, 1992; Squire et al., 1992), in a region that may be a homologue to theoccipital-temporal “what” pathway in nonhuman primate vision (as opposed to the occipital-parietal“where” or “action” pathway; Felleman & Van Essen, 1991; Goodale & Milner, 1992; Haxby et al., 1991;Ungerleider & Mishkin, 1982). These areas accept retinotopically coded inputs from primary visualcortex (Fox, Miezin, Allman, Van Essen, & Raichle, 1987; Kosslyn et al., 1993; Tootell, Silverman, Switkes,& De Valois, 1982), and they appear to send output representations that signal a recognized form tonon-visual subsystems (e.g., phonological, conceptual/associative, motoric, etc.) as well as to other visualsubsystems (e.g., the occipital-parietal “where” or “action” pathway).

Computational ConstraintsCareful considerations of our visual abilities, in terms of goals and strategies, suggest that these

visual-form areas do not perform a single or simple process. Given retinotopically coded images offorms that appear in the world, visual-form subsystems accomplish at least two essential goals. First, theyunderlie the ability to recognize abstract categories of forms. For example, when reading a book, areader usually categorizes word forms at only the coarse level of classification needed to access theappropriate meanings associated with the words. More concretely, one categorizes forms at the coarselevel in which all of the forms in Figure 1 belong to the same category and hence produce the sameoutput. However, visual subsystems also underlie another important ability, that of recognizing specificinstances within the same abstract category of form. For example, to recognize a signature, one usuallycategorizes a letter string at the fine-grained level needed to distinguish a restricted set of the manypossible ways in which the same letter string can appear. More concretely, one categorizes forms at thefine-grained level in which the forms in Figure 1 produce different outputs.

---------------------------------Insert Figure 1 about here

---------------------------------According to many theories, these two abilities should be accomplished in a single,

undifferentiated processing subsystem (for general computational theories, see Knapp & Anderson, 1984;McClelland & Rumelhart, 1985; for object recognition theories, see Hummel & Biederman, 1992; Tarr,1995; see also most theories of word recognition). However, the two abilities seem to place contradictorydemands on the relevant processing mechanisms. To recognize abstract categories, it should be usefulfor a subsystem to ignore the visually distinctive information that differentiates specific instances in acategory and to focus on the information that is relatively invariant across instances. For example, onlythe shared information between the two input forms in Figure 1 is very useful for accomplishing the goalof recognizing the common abstract category. Note that the relatively invariant information across theforms in Figure 1 is only found in parts of the larger wholes (see top of Figure 2). In contrast, torecognize specific instances, it is necessary for a subsystem to focus on just the sort of information thatmay be effectively ignored when recognizing abstract categories. For example, the information thatdistinguishes the inputs in Figure 1 must be processed to accomplish the goal of discriminating theforms. It is important to note that the visually distinctive information that differentiates the forms inFigure 1 in addition to the specific instances of other words is found in the wholes of those forms (seebottom of Figure 2).

---------------------------------Insert Figure 2 about here

---------------------------------


Implementational ConstraintsIn addition, considerations of findings in the neuropsychological literature suggest that there are

important differences in how visual-form subsystems operate across the different physical substrates ofthe left hemisphere (LH) and right hemisphere (RH). First, subsystems in the LH and RH play importantroles in word recognition and face recognition, respectively (e.g., Damasio & Damasio, 1983; Geffen,Bradshaw, & Wallace, 1971; Petersen & Fiez, 1993; Rhodes, 1985; Sergent et al., 1992; Sergent & Bindra,1981). Second, recognition of word, face, and object forms appears to be performed through two, notone or three, independent capacities. In a review of the visual associative agnosia literature, Farah(1990, 1991) noted that only a subset of the logically possible combinations of impaired abilities exhibitedby individual brain-damaged patients are found frequently. The combinations indicate that two kinds ofvisual recognition capacities are susceptible to damage, one involving word and sometimes objectrecognition (and affected by LH damage), and the other involving face and sometimes object recognition(and affected by RH damage). These results have direct implications for abstract-category and specific-instance recognition, to the extent that word recognition usually relies on abstract-category processing,face recognition usually relies on specific-instance processing, and object recognition likely relies ondifferent processes (abstract versus specific) in different circumstances.

Relatively Independent Visual-Form SubsystemsAlthough these considerations do not rule out the possibility that a single subsystem underlies

both abilities, they do lead to the theory that relatively independent subsystems underlie visual-formrecognition. We have hypothesized that an abstract visual-form (AVF) subsystem underlies recognition ofabstract categories of forms, processes relatively-invariant input information through the use of a parts-based internal processing strategy, and operates more effectively in the LH than in the RH. In contrast, aspecific visual-form (SVF) subsystem underlies recognition of specific instances of forms, processesvisually-distinctive input information through the use of a holistic internal processing strategy, andoperates more effectively in the RH than in the LH. These subsystems may operate relativelyindependently in large part because they rely on contradictory internal processing strategies. In the nextsection, we summarize initial behavioral evidence in support of this theory. Afterward, we usecomputational analyses and modeling to suggest a concrete specification of these internal processingstrategies and to illuminate their contradictory natures.

BEHAVIORAL EVIDENCE FOR RELATIVELY INDEPENDENT SUBSYSTEMSWe have tested the AVF/SVF subsystems theory in the following studies. The initial behavioral

tests utilized divided-visual-field presentations of visual forms during the test phases of various memoryexperiments.

Rationale for Divided-Visual-Field StudiesWe use the divided-visual-field technique as a tool to help test whether a single visual-form

subsystem operates in a fairly unitary manner in the brain or two subsystems operate in a relativelyindependent manner. Our rationale is similar to the rationale that interactions between task and field ofvisual presentation are needed to adequately study laterality effects in this paradigm (Hellige, 1983). Ofcourse, in a divided-visual-field presentation, the information presented directly to one hemisphere mustcross brain commissures to be processed by the other. The important implication is that mechanisms inthe first hemisphere obtain higher quality information (e.g., Dimond, Gibson, & Gazzaniga, 1972; Gross,Rocha-Miranda, & Bender, 1972) and obtain it more quickly than mechanisms in the other hemisphere.Thus, if the characteristic processing of one hypothesized subsystem (e.g., AVF processing) is performed


more effectively when high quality visual input is processed initially in one hemisphere (e.g., left) than inthe other, whereas the characteristic processing of a different hypothesized subsystem (e.g., SVFprocessing) is performed more effectively when high quality visual input is processed initially in the otherhemisphere (e.g., right) compared with the one yielding the first advantage, then two sorts of processingmust rely on two sets of neural circuitry that operate at least relatively independently.

Note that, taken alone, such behavioral results do not suffice to indicate the extent to which thesubsystems are lateralized or the degree to which they are modular. They only indicate that at leastrelatively independent subsystems are involved and that they are at least weakly lateralized. Furtherinvestigations utilizing additional methodologies are needed to clarify the unresolved issues. Forexample, further methods are needed to determine whether one subsystem operates in only onehemisphere, whether both subsystems operate with asymmetric efficiency in each hemisphere, etc.

Visual Repetition PrimingIn one study, we examined repetition priming for visual word forms (Marsolek, Kosslyn, &

Squire, 1992). Subjects were asked to read lists of common words presented in the central visual fieldduring an initial encoding phase. Half of the words were presented in all lowercase letters (e.g.,“convince”), and half were presented in all uppercase letters (e.g., “PRIMARY”). Afterward, in apresumably unrelated second phase of the experiment, subjects completed word stems (three-letterbeginnings of words that can be completed to form many common words; e.g., “con”) to form the firstwords that came to mind. Each stem was presented in the left or right visual field. Repetition primingwas exhibited when they produced word completions that corresponded to words that were viewedearlier in the experiment with a greater-than-chance tendency. The results were that such priming wasgreater when stems were presented in the same letter case as previously presented words, compared withthe different letter case. More important, this letter-case-specific priming effect was found when thestems were presented directly to the RH (briefly in the left visual field) but not when they were presenteddirectly to the LH (briefly in the right visual field). In a related study, Marsolek, Squire, Kosslyn, &Lulenski (1994) discovered that, in certain experimental conditions, the same pattern of results isobtained when subjects use the stems as cues to help them explicitly recall previously seen words.

These results suggest that the structural changes that underlie visual memory effects may bequalitatively different across subsystems that operate asymmetrically in the two hemispheres. Structuralchanges that underlie storage of the visually distinctive information that differentiates specific instancesin an abstract category of form (e.g., lower- vs. uppercase versions of the same word) are instantiatedmore effectively in the RH than in the LH. Given that this sort of information storage should becharacteristic of an SVF subsystem, the results indicate that an SVF subsystem, but not an AVF subsystem,operates more effectively in the RH than in the LH.

Task Demands in Visual Repetition PrimingIn a similar study, we examined how task demands may influence priming in visual-form

subsystems (Burgund & Marsolek, 1996). All subjects read lists of common words and pronounceablenonwords (half of each in all lowercase letters, and half of each in all uppercase letters) that wereintermixed and presented in the central visual field during initial encoding. Then, in the test phase ofone experiment, subjects performed a standard perceptual identification task by identifying brieflypresented letter strings and writing them down, without following any particular instructions on how towrite the strings. In the test phase of a different experiment, subjects performed a form-specificperceptual identification task by identifying briefly presented letter strings and writing them down in thesame letter case as they had appeared on the computer monitor. Repetition priming was measured as the


tendency to identify and report letter strings that had been viewed earlier in the experiment moreaccurately than letter stings that had not been processed previously. The results were that letter-case-specific priming was greater when test items were presented directly to the RH than when they werepresented directly to the LH (as observed in word-stem completion studies; Marsolek et al., 1992, 1994),but this was true only when the form-specific perceptual identification task was performed and not whenthe standard perceptual identification task was performed (cf. Koivisto, 1995).

These results suggest that task demands directly affect which subsystems are recruited in differentpriming tests. Performance in the standard perceptual identification task may be influenced by processingin an AVF subsystem to a greater degree than by processing in other subsystems, because letter-case-specificinformation does not need to be processed for accurate performance (unlike in the form-specificperceptual identification task). In addition, there is only one correct response per trial in this task, whichmay have the effect that the most efficient subsystem for the job (i.e., an AVF subsystem) dominates theproduction of a response (unlike in the word-stem completion task, in which the large number of “correct”responses per stem may have the effect that various subsystems contribute to the production of a response).Of course, performance in the form-specific perceptual identification task should be influenced highly by anSVF subsystem, because letter-case-specific information must be processed for accurate performance. Inlight of these task demands, the results further support the AVF/SVF subsystems theory.

Visual ClassificationIn another experiment, we investigated classification of novel visual forms (Marsolek, 1995).

Subjects first learned to associate labels to categories of unfamiliar letterlike forms, and each form waspresented in the central visual field during learning. Afterward, subjects were asked to classify test formsusing the newly-learned categories. The results were that they classified the previously unseenprototypes of the newly learned categories (each prototype was the central tendency of the instances inone category) more effectively when they were presented directly to the LH than to the RH. In contrast,subjects classified the previously seen specific instances more effectively when they were presented directlyto the RH than to the LH (and they did not classify previously unseen non-prototype instances differentlydepending on hemisphere of presentation).

Generally, the prototypes of these visual-form categories possessed a large amount of the visualinformation that remained relatively invariant across the different instances in one category. In fact, eachprototype contained a larger amount of this relatively invariant information than did any of the otherspecific instances in its category. Hence, the results indicate that the information that is useful for an AVFsubsystem (relatively invariant information) is stored more effectively in the LH than in the RH. Incontrast, the previously seen specific instances of course contained visually distinctive information.Hence, the results indicate that the information that is useful for an SVF subsystem is stored moreeffectively in the RH than in the LH. Note that a RH advantage for processing the previously seen specificinstances was obtained even though the demands of the categorization test task may have favoredprocessing in an AVF subsystem more than processing in an SVF subsystem. We suggest that an SVFsubsystem contributed to performance in this task nonetheless, because all of the forms were unfamiliarpreexperimentally, and an SVF subsystem should store novel visual forms especially well. This reasoningfor this hypothesis (see below) stems from our analyses of the internal processing strategies for AVF andSVF subsystems.

CONTRADICTORY INTERNAL PROCESSING STRATEGIESMany previous theories of functional asymmetries have emphasized distinctions that can be

understood in terms of analytic/parts-based processing versus holistic processing in the left and right


hemispheres, respectively (e.g., Bever, 1980; Bradshaw & Nettleton, 1981; Corballis, 1989; Diamond &Carey, 1986; Farah, 1990, 1991; Levine & Calvanio, 1989). Unfortunately, little attention has been given toexplicating precisely the differences between these two kinds of processes (see Marshall, 1981, forimportant problems with analytic vs. holistic distinctions). In addition, little attention has been given toclarifying the conditions under which each kind of processing should be recruited or on explaining whythe two kinds of processing appear to operate in different parts of the brain.

In this section, we offer computational analyses suggesting that one visual-form subsystem (anAVF subsystem) performs parts-based processing, because that sort of processing is required for effectiverecognition of abstract categories per se. Another visual-form subsystem (an SVF subsystem) performsholistic processing, because that sort of processing is required for effective recognition of specificinstances per se. This analysis illuminates a concrete specification of the distinction between--and thecontradictory natures of--parts-based versus holistic internal processing. In addition, we discuss howresults from a neural network modeling study supplement and directly test this computational reasoning.

Neural Network Models of Visual-Form RecognitionWe have examined relatively simple feedforward neural network models that receive

“retinotopically-coded” bit-mapped inputs, in an effort to investigate how parallel distributed processingsystems might accomplish the task of recognizing such input forms (Marsolek, 1994). Each network wascomposed of three layers of processing units: an input layer, a hidden layer, and an output layer. Inputunits were connected to hidden units via weighted interconnections, and in turn hidden units wereconnected to output units via weighted interconnections. In each processing trial, an input form waspresented to the network as a pattern of activation across the two-dimensional array of input units.These input forms were just smaller than the array of input units, and in one trial an input form waspresented in either the upper-right, lower-right, lower-left, or upper-left region of the array. After aninput form was presented, activation flowed across the first set of weighted connections to activate thehidden units, and in turn activation flowed across the second set of weighted connections to activate theoutput units. Weights on the internal connections modulated the flow of activation between layers.

Each network was trained to perform input-to-output mappings through the use of an error-correction procedure (backpropagation-of-error; Rumelhart, Hinton, & Williams, 1986). This sort oftraining was used to guide each network to discover a set of weights across its internal connections thatallowed the intended input-output mappings to take place. When such networks are trained toaccomplish input-output mappings that apparently take place in certain areas of visual cortex, they tendto discover mapping solutions that the relevant neural subsystems appear to use (e.g., Churchland &Sejnowski, 1992; Lehky & Sejnowski, 1988; O'Reilly, Kosslyn, Marsolek, & Chabris, 1990; Zipser &Andersen, 1988). Note that any neural implausibility of the specific backpropagation training processused in this study was not important, in part because networks that use training algorithms that are verysimilar to backpropagation, yet are biologically plausible, tend to produce results that are highly similarto those found with backpropagation (see, e.g., Mazzoni, Andersen, & Jordan, 1991).

The input patterns and the target output patterns in our networks were random-dot patternsthat adhered to statistical similarity constraints. However, in the following discussion we describe themodels as though word forms like “BEAR,” “bear,” “bear,” “bear,” etc., served as input patterns, simplyfor clarity.

AVF, SVF, and Intermediate Input-Output MappingsEach network was trained to solve two of the following three input-to-output mapping problems,

with the three mapping problems examined in different pairs across different networks. In AVF


mappings (see Figure 3), the uncorrelated input patterns “BEAR” and “bear” (note the visualdissimilarities between these forms), in addition to the correlated input patterns “bear” and “bear” (notethe visual similarities between these forms), were all mapped to the same output representation.However, in SVF mappings (see Figure 4), each specific instance (e.g., “BEAR,” “bear,” “bear,” and “bear,”etc.) was mapped to a different output representation. Furthermore, in an intermediate-level mappingproblem (see Figure 5), the correlated input patterns “BEAR” and “BEAR” were mapped to the sameoutput representation, yet the input patterns “bear” and “bear,” which were correlated with each otherbut not with “BEAR” and “BEAR,” were mapped to a different output representation. Note that the inputpatterns were the same for all three mapping problems; only the mappings to output representationsdiffered across the three mappings.

---------------------------------Insert Figures 3, 4 and 5 about here

---------------------------------We examined whether the internal architectures of the networks affected their abilities to

perform different pairs of mapping problems. One set of networks was trained to perform both AVF andSVF mappings of input patterns through different subsets of output units, half of the output unitsallocated to the AVF task and the other half to the SVF task. Some of these networks were unified or“unsplit” models, in which all hidden units were connected to all output units. However, other networkswere “split” (e.g., Rueckl, Cave, & Kosslyn, 1989), in that one subset of the hidden units was connectedonly to the AVF output units and the other subset was connected only to the SVF output units. In thisway, split networks had separate subcomponents devoted to the different mapping tasks, but unsplitnetworks accomplished the two tasks through a unified model.

Results were that, after an arbitrary number of training trials, models with separatesubcomponents performed the AVF and SVF mappings more efficiently than unified models. (These splitnetworks also outperformed networks with the same number of hidden-to-output unit connections as thesplit networks but with no systematic splitting of the hidden units into separate pools allocated to thedifferent mappings.) It is important to note that this separate-subnetwork advantage did not extend tomodels that performed both AVF and intermediate-level mappings. Furthermore, the separate-subnetwork advantage did not extend to models that performed both SVF and intermediate-levelmappings (see also Knapp & Anderson, 1984; McClelland & Rumelhart, 1985). Hence, split networks didnot always outperform unsplit networks; they did so in this study only when AVF and SVF mappings, inparticular, were performed in the same networks.

The split networks described above were strongly split networks in that they contained a strongmodularity in their internal architectures. Strong modularity was not necessary, however, for theseparate-subnetwork advantage. AVF and SVF networks that were weakly split (such that one subset ofhidden units was dedicated to the AVF task, another to the SVF task, and a third to both tasks) alsooutperformed unsplit networks, even though weakly-split AVF and intermediate networks and weaklysplit SVF and intermediate networks did not outperform their unsplit counterparts.

The idea that SVF and intermediate mappings are compatible with one another in this study(Marsolek, 1994), as well as in other computational studies (see Knapp & Anderson, 1984; McClelland &Rumelhart, 1985), is one aspect of the present theory and models that is different from Kosslyn's (1987;Kosslyn, Chabris, Marsolek, & Koenig, 1992) theoretical distinction between coordinate and categoricalspatial relations encoding and from Jacobs and Kosslyn's (1994) neural network models of “coordinateshape” and “categorical shape” processing. Their coordinate-categorical mapping distinction is verysimilar to our SVF-intermediate mapping distinction, in that one of the tasks (SVF or coordinate) involvesmapping visually similar inputs to different output representations and the other task (intermediate or


categorical) involves mapping visually similar inputs, and only inputs that are similar to them, to the sameoutput representation. Yet, we find that SVF and intermediate mappings are compatible, whereas Jacobsand Kosslyn (1994) conclude that different mapping solutions are useful for coordinate-shape andcategorical-shape processing in their neural network models.

The apparently inconsistent results across these modeling studies may be resolved with the helpof additional analyses (Marsolek, 1994). In our study, the projective fields of the hidden units in all splitnetworks were held to a constant size across the pools of AVF, SVF, and intermediate-level hidden units.The projective field of a hidden unit is determined by its connections to the output units (Lehky &Sejnowski, 1988), thus split networks with a larger number of output units devoted to the SVF mappingthan to the intermediate mapping would have SVF hidden units with larger projective fields than thoseof the intermediate-mapping hidden units. This is important because methodological difficulties arisewhen the two subsets of hidden units in a split network do not have the same-sized projective fields,which was the procedure used by Jacobs and Kosslyn (1994). When SVF hidden units have largerprojective fields than intermediate-level hidden units, we have found that split networks outperformtheir unsplit counterparts, in restricted conditions, but the result is equivocal (Marsolek, 1994). The splitadvantage in this situation could be due to facilitation from splitting networks into large and smallprojective-field subnetworks (which may have to do with optimal learning rates being proportional tofan-in and fan-out of processing units in networks like these; cf. Plaut & Hinton, 1987), regardless ofwhether contradictory mapping tasks are performed in the separate subnetworks.

Therefore, we conclude that AVF and SVF mappings, but not the other pairs of mappings, areperformed more efficiently through separate subnetworks than through unified networks. These resultsindicate that the internal processing strategies that are useful for parallel distributed processing systemsto perform AVF mappings and the internal processing strategies that are useful for such systems toperform SVF mappings may interfere with one another or are contradictory in some way. In what waysare the strategies contradictory? Our computational argument is that different internal processingstrategies are useful for subsystems that perform AVF and SVF recognition. Parts-based processing shouldbe useful for an AVF subsystem because the relatively invariant information of the forms in one categorytend to be found in their parts (see Figure 2). In contrast, holistic processing should be useful for anSVF subsystem because the visually distinctive information that differentiates specific instances of forms isfound in the holistic structures of the forms (see Figure 2). A single, undifferentiated mechanism cannotperform both parts-based and holistic processing effectively. For proper assessment, this hypothesisrequires the following explication.

AssumptionsTwo important assumptions are needed for the computational analysis offered here. First,

visual-form subsystems (and artificial networks that simulate them) receive the retinotopically-mappedrepresentations of input forms that are currently being captured by selective attention mechanisms. Thatis, visual selective attention serves to filter out extraneous inputs and acts to “surround” the form thatcurrently is being processed, regardless of the location or size of the form as it appears on the retina(e.g., LaBerge, 1995). Such a selective-attention filter may operate at least in part before processing invisual-form subsystems of the inferior temporal cortex, because cells in the inferior temporal region donot change their response selectivities appreciably when visual forms change locations and sizes yetremain within the receptive fields of the cells (Schwartz, Desimone, Albright, & Gross, 1983). Thisassumption is needed to account for effects of size and location invariance in visual-form priming;changes in size and location of forms between initial encoding and subsequent test do not greatlyinfluence priming effects (Biederman & Cooper, 1992; Cooper, Schacter, Ballesteros, & Moore, 1992).


The analogy in the network models is that input forms are conceptualized to be scaled so that they arejust smaller than the input grid in any one trial, regardless of their sizes or locations on the “retina” insome earlier stage of processing.

The second important assumption is that visual-form priming is produced by structural changesin visual-form subsystems (and in artificial networks that simulate them). This assumption is needed toaccount for relatively long-term priming effects, which can last for several days in the case of objectpriming, even for amnesic patients (Cave & Squire, 1992). Long-term priming must be supported throughsome kind of physical change that serves to store information for a period of time (possibly through localsynaptic changes such as in long-term potentiation in visual cortex; Artola & Singer, 1987; Komatsu, Fujii,Maeda, Sakaguchi, & Toyama, 1988). The analogy in the network models is that priming is supported bysmall changes in the previously established weights on the internal connections of the networks due torecent processing of a prime stimulus (and the accompanying backpropagation-of-error). Different kindsof weight changes can be considered in these network models, depending on the kind of informationthey represent.

Parts-Based Internal ProcessingImagine the receptive fields that would develop in these network models during training (the

receptive field of a hidden unit is determined by its connections to the input units). If each hidden unitin a network were to have a relatively large absolute weight value on only a few of the connectionsfeeding into it and very small weight values (perhaps 0) on the other connections, then each hidden unitwould be connected functionally to only a few input units. If different hidden units were sensitive todifferent subsections of the input array, such that any one hidden unit were to be sensitive to only a partof any one input form, this would be an example of a network with a parts-based internal representationstrategy. Such a network may utilize what is called a “sparse” or “local” coding strategy (see Churchland& Sejnowski, 1992). In this case, different hidden units necessarily represent different parts of any oneinput form.

Note that this kind of internal processing strategy should be useful for an AVF subsystem. Forexample, an efficient internal representation for the abstract category in Figure 1 may be activation of thehidden units that are sensitive to that category's relatively invariant information (see Figure 2), coupledwith little or no activation of the hidden units that are sensitive to the parts that are found in the visuallydistinctive information that differentiates specific instances in that abstract category or to the parts thatare found in other abstract categories. Indeed, examinations of the trained networks described above(Marsolek, 1994) indicate that this sort of internal processing strategy is used in the AVF portions of splitnetworks. The hidden units in these AVF subnetworks develop receptive fields that utilize a relativelyparts-based strategy.

The priming in such a network necessarily would be parts-based. Because the activation of anyone hidden unit represents only one part of an input form, small weight changes on the connectionsfeeding into different hidden units would yield priming for relatively independent information aboutdifferent parts of an input form. Hence, different parts of the same input form should be primedrelatively independently in such a subsystem. Indeed, behavioral results summarized below support thishypothesis.

Holistic Internal ProcessingNow imagine another possibility in these network models. If each hidden unit in a network were

to have relatively large (but varying across units) absolute weight values on perhaps all of the connectionsfeeding into it, then each hidden unit would be connected functionally to the whole (or at least a very


large portion) of the input array. If different units were differentially sensitive to different parts of thewhole input array, such that different hidden units were sensitive to slightly different aspects of thewhole of any one input form, this would be an example of a network with a holistic internalrepresentation strategy. (For advantages in using “coarsely” coded representations like these, seeBallard, 1986; Hinton, McClelland, & Rumelhart, 1986.) In this case, the hidden units would notrepresent different parts of an input form explicitly as such, only implicitly as portions of the whole formto which perhaps every unit is sensitive.

This kind of internal processing strategy should be useful for an SVF subsystem. For example, anefficient internal representation for the specific instance “BEAR” may be a distinct pattern of activationacross perhaps all of the hidden units, each of which may be slightly differentially sensitive to all of theinformation in that particular input form. Indeed, examinations of the trained networks described above(Marsolek, 1994) indicate that this sort of internal processing strategy is used in the SVF portions of splitnetworks. The hidden units in these SVF subnetworks develop receptive fields that utilize a relativelyholistic strategy.

The priming supported by such a network would be holistic. Because the activation of any onehidden unit represents holistic structure, small weight changes on the connections feeding into any onehidden unit would yield priming for the wholes of forms. Hence, different parts of the same input formcould not be primed independently in such a subsystem. Indeed, behavioral results summarized belowsupport this hypothesis.

In addition, it is worthwhile to note that the holistic priming supported by such a networkshould be useful for priming of unfamiliar forms. At some level, every unfamiliar form has parts (in thelimit, various sorts of edges) that are familiar; it is the holistic structure of any unfamiliar form thatcontains the information that makes it unfamiliar. Hence, the novel aspects of unfamiliar forms shouldbe primed more effectively in a holistic processing network than in a parts-based processing network.This hypothesis is also supported by behavioral results summarized below.

Contradictory StrategiesIt should be clear that a single, undifferentiated network could not implement both a parts-based

and a holistic processing strategy. The same mechanism could not represent parts explicitly as such andnot represent parts explicitly as such. Although the parts-based and holistic processing strategiesdescribed above are relatively extreme versions of both, less extreme versions of the two strategies alsomay be contradictory. The relatively invariant information associated with an abstract category like inFigure 1 is present in some subset of the information in any one input form, a subset that necessarily issmaller than the amount of visually distinctive information that differentiates any one specific instance inthat category (i.e., the holistic structure of that instance). Likewise, the visually distinctive informationthat differentiates specific instances in Figure 1 is present in the holistic information in these inputs, a setof information that necessarily is larger than the relatively invariant information for that abstract category(i.e., parts of the forms). Hence, efficient internal representations for AVF categorizations should involveparts-based coding of a relatively small amount of information per form, whereas efficient internalrepresentations for SVF categorizations should involve holistic coding of a relatively large amount ofinformation per form. It would be difficult for a single mechanism to store both kinds of informationeffectively (only the relatively invariant information for the abstract category as well as the visuallydistinctive information that differentiates specific instances in that category). In other words, it would bedifficult for a single mechanism to accomplish both AVF and SVF recognition effectively.

Given that contradictory strategies (parts-based vs. holistic) are useful for accomplishing differentgoals (AVF vs. SVF categorizations), and both goals are vitally important for survival, selective pressures


may have led to the evolution of relatively independent processing subsystems with the differentsubsystems utilizing different computational strategies. Note that the claim here is not that it would beimpossible for a single subsystem to accomplish both AVF/parts-based and SVF/holistic processes or thatthese processes are incompatible per se. Instead, the reasoning involves considerations of relativeefficiency and of course should be tested. Next, we summarize behavioral evidence in support of ourcomputational theory.

BEHAVIORAL EVIDENCE FOR PARTS-BASED VERSUS HOLISTIC PROCESSINGIn further behavioral studies, we have tested directly whether an AVF subsystem utilizes a parts-

based internal processing strategy whereas an SVF subsystem utilizes a holistic processing strategy. Recentbehavioral studies using divided-visual-field presentations of memory test items and studies using aninterhemispheric communication paradigm support these hypotheses.

Visual Repetition PrimingThe computational theory is supported by a recent study of visual priming using the word-stem

completion task (Marsolek et al., 1995). In this study, some of the word stems (and hence the beginningsof their corresponding words) were composed of letters with visually dissimilar lower- and uppercasestructures (e.g., “bea” / “BEA”), whereas the other items were composed of letters with visually similarlower- and uppercase structures (e.g., “sco” / “SCO”). Like the priming studies described above (Marsoleket al., 1992; 1994), subjects read lists of words presented in the central visual field during an initialencoding phase. Half of the words were presented in all lowercase letters, and half were presented inall uppercase letters. Then, subjects completed word stems to form the first words that came to mindduring a presumably unrelated second phase of the experiment.

Results indicate that the parts-based information that is common to “BEAR” and “bear” is storedin an AVF subsystem that operates more effectively in the LH than in the RH. Same and different letter-case priming did not differ when the dissimilar stems were presented directly to the LH (even thoughsame-case priming was greater than different-case priming when these stems were presented directly tothe RH). For example, “BEAR” primed “bea” as well as “bear” primed “bea” in LH stem presentations.(Note that this effect apparently is supported by visual subsystems per se, because “BEAR” primes “bear”to a greater degree than hearing the word bear primes the visual form “bear” [Bowers, in press].) Incontrast, the holistic information that differentiates even structurally similar forms, like “SCOOP” and“scoop,” is stored in an SVF subsystem that operates more effectively in the RH than in the LH, albeitthrough interactions with the hippocampal formation (see several chapters in Schacter & Tulving, 1994;Cohen & Eichenbaum, 1993; McClelland, McNaughton, & O'Reilly, 1995; Squire, 1992). This qualificationis needed because only explicit memory as measured in word-stem cued recall, but not repetition primingas measured in word-stem completion priming, produces greater same-case than different-case memoryfor the similar-case items in RH presentations. For example, when “scoop” had been presented earlier inthe experiment and “sco” was the test cue, that word was recalled more readily than when “SCOOP” hadbeen presented earlier and “sco” was the test cue, in RH but not in LH presentations of the cues.

Visual Priming for New AssociationsIn further experiments, we examined visual priming for new associations between previously

unrelated words (Marsolek, Schacter, & Nicholas, 1996). During the encoding phase, subjects read listsof word pairs, one word presented above the other in the central visual field for each pair. Half of thepairs were presented in all lowercase letters, and half were presented in all uppercase letters. Afterward,subjects completed word stems that were presented beneath complete context words in a presumably


unrelated test phase. Letter-case-specific priming in stem completion was found only when the contextwords were the same words that had appeared previously above the primed completion words duringinitial encoding and when the two items in a test pair (context word and word stem) were presenteddirectly to the RH. When the context words were different words from those which had appeared abovethe primed completion words during initial encoding, no letter-case-specific priming was obtained instem completion. These results indicate that priming for novel holistic information (i.e., one word formas it appears above another word form) is supported by a subsystem that distinguishes lower- versusuppercase versions of the same word and operates more effectively in the RH than in the LH, as predictedfrom the computational theory described above.

Interhemispheric Communication of Visual-Form InformationIn another behavioral study, we examined interhemispheric transfer of visual-form information

(Nicholas & Marsolek, 1996). Subjects were asked to compare two visually presented letters in each trial.The two letters appeared in the same visual field in half of the trials, but in different visual fields in theother half of the trials, similar to a task used by Banich & Belger (1990) who presented three letters pertrial and asked subjects to compare the bottom-most letter to the other two in each trial. With such aprocedure, interhemispheric transfer of information must have taken place when the comparison letterswere presented across hemispheres (briefly displayed in different visual fields) but not necessarily whenthey were presented within a hemisphere (briefly displayed in the same visual field). We took advantageof this circumstance in experiments investigating two visual comparison tasks. Results indicate that AVFand SVF subsystems are differentially affected by interhemispheric transfer of visual-form information inways that are predictable from the computational theory described above.

In an AVF comparison task, subjects decided whether the pairs corresponded to the same letterof the alphabet (e.g., “a” and “A,” “s” and “S”) or not (e.g., “a” and “Q”, “s” and “P”). In this task, theyperformed more accurately in across-hemisphere trials than in within-hemisphere trials. However, thisacross-hemisphere advantage was found for similar-case letters (e.g., s/S), but not for dissimilar-caseletters (e.g., a/A) which yielded no within- or across-hemisphere advantage. Current theories ofinterhemispheric communication (Banich & Belger, 1990; Belger & Banich, 1992; Hellige, 1993b) wouldnot predict this finding, yet the AVF/SVF subsystems theory may account for it. The effects of noiseproduced by interhemispheric transfer of similar-case letters may not be as detrimental as the effects ofnoise produced by interhemispheric transfer of dissimilar-case letters, in the AVF task. That is, noisyversions of the relatively invariant information in similar-case letters may be processed by an AVFsubsystem more effectively than noisy versions of the relatively invariant information in dissimilar-caseletters, because there is more relatively invariant information per letter for similar-case items than fordissimilar-case items to help overcome the noise. Hence, the relatively invariant information needed tomake AVF categorizations of similar-case letters, but not dissimilar-case letters, may cross braincommissures effectively enough to take advantage of the benefits of distributing the processing of the twoletters across the two hemispheres (cf. Banich, 1995; Dimond & Beumont, 1971).

In an SVF comparison task, subjects decided whether the letters in a pair were physically thesame (e.g., “a” and “a,” “S” and “S”) or not (e.g., “a” and “A,” “s” and “S”). In this task, they performedmore accurately in within-hemisphere trials than in across-hemisphere trials. Hence, SVF processingapparently cannot take advantage of the benefits of distributing the processing of the two letters acrossthe two hemispheres. Interestingly, this within-hemisphere advantage was found for similar-case letters(e.g., s/S), but not for dissimilar-case letters (e.g., a/A) which yielded no within- or across-hemisphereadvantage. Current theories of interhemispheric communication (Banich & Belger, 1990; Belger &Banich, 1992; Hellige, 1993b) would not predict this result, but the AVF/SVF subsystems theory would.


The visually distinctive information needed to make SVF categorizations of dissimilar-case letters may notcross brain commissures (to take advantage of neural distribution) effectively enough to produce anacross-hemisphere advantage. Moreover, the visually distinctive information needed to make SVFcategorizations of similar-case letters may be so fine-grained that it crosses brain commissures soineffectively that a within-hemisphere advantage is produced. Indeed, by hypothesis, an SVF subsystemshould be fairly sensitive to the noise produced by interhemispheric transfer of information, because itmay treat noise as the kind of “visually distinctive” information that it is tuned to process when it is calledupon to perform SVF categorizations.

CONCLUSIONS AND IMPLICATIONSA fundamental aspect of the architecture of our neural processing subsystems appears to be that

relatively independent subsystems underlie AVF versus SVF recognition. In this chapter, we havesummarized research that supports this conclusion through an integration of behavioral studies andcomputational analyses and models. We have highlighted considerations of the inputs available to visualprocessing subsystems, the goals that these subsystems must satisfy to underlie important visual abilities,and plausible internal processing strategies that these subsystems could use to achieve the appropriategoals given the available input. All of these considerations are combined to characterize explicitly thefunctions--and contradictory natures of--relatively independent AVF and SVF subsystems. We conclude thatan AVF subsystem operates more effectively in the LH than in the RH, and it uses a parts-based internalprocessing strategy to focus on the relatively invariant information in visual-form inputs, which allows it torecognize abstract categories of visual forms effectively. In contrast, an SVF subsystem operates moreeffectively in the RH than in the LH, and it uses a holistic internal processing strategy to capture thevisually distinctive information in visual-form inputs, which allows it to recognize specific instances offorms effectively.

One set of results from previous research that we have not mentioned heretofore is consistentwith our conclusions and may help to illuminate an important feature of our theoretical approach. Individed-visual-field studies, subjects identify and discriminate high spatial-frequency information moreeffectively when it is presented directly to the LH than to the RH, whereas they identify and discriminatelow spatial-frequency information more effectively when it is presented directly to the RH than to the LH(e.g., Christman, Kitterle, & Hellige, 1991; Kitterle, Christman, & Hellige, 1990; Kitterle & Selig, 1991). Ofcourse, these asymmetries may help to explain why LH advantages are found when subjects process thelocal parts of hierarchically arranged stimuli whereas RH advantages are found when they process theglobal forms of these stimuli (Robertson & Lamb, 1991; Van Kleeck, 1989), given that relatively highspatial-frequency information should be useful for processing the local forms and relatively low spatial-frequency information should be useful for processing the global forms (Sergent, 1982).

These findings are consistent with our theory for the following reasons. A subsystem that hasevolved to perform AVF recognition should perform parts-based processing effectively. An individual cellin such a subsystem should receive information from a relatively small portion of the input array, andhence the cell optimally should be tuned to a high spatial-frequency band. It is important to note thatthe cell also may be sensitive to even higher frequencies in some conditions, but it may not be verysensitive to substantially lower frequencies (e.g., Sergent, 1989). It is also important to note that acollection of such cells, each receiving information from different but overlapping portions of the input,should be very sensitive to even higher frequencies than an individual cell’s optimal frequency, throughthe use of a distributed representation (cf. Hinton et al., 1986). Furthermore, this collection of cells maybe able to respond to relatively low frequencies as well, given that different cells in the distributedrepresentation should respond differently due to the different levels of overall activation that are presentin their different portions of a low spatial-frequency input.


In contrast, a subsystem that has evolved to perform SVF recognition should perform holisticprocessing effectively. An individual cell in such a subsystem should receive information from a relativelylarge portion of the input array, and hence it optimally should be tuned to a low spatial-frequency band.Of course, by the same analysis as above, the cell also may be sensitive to higher frequencies in someconditions, but not to substantially lower frequencies (e.g., Sergent, 1989). Furthermore, a collection ofsuch cells, each receiving slightly different information from large portions of the input, should besensitive to higher frequency information than that to which an individual cell in the subsystem issensitive (cf. Hinton et al., 1986), and the collection of cells should be sensitive to lower frequencies aswell.

Thus, both subsystems may process high and low spatial-frequency information, but not with thesame efficacy. Assuming that the optimal frequency band for most individual cells in a subsystemdetermines the optimal range of frequencies for distributed representations in that collection of cells,and assuming roughly the same number of cells per subsystem, we conclude the following. Comparedagainst one another, an AVF subsystem should process relatively high spatial-frequency information moreeffectively than an SVF subsystem, and an SVF subsystem should process relatively low spatial-frequencyinformation more effectively than an AVF subsystem.

Of course, the more general spatial-frequency hypothesis can account for these hemisphericasymmetries as well. The LH may be specialized for processing high spatial-frequency informationwhereas the RH may be specialized for processing low spatial-frequency information (Sergent, 1982).This hypothesis certainly has been useful for attracting attention to input factors and their roles inhemispheric asymmetries. However, concentration on input factors without their relations to otherfactors leaves an incomplete picture of hemispheric asymmetries.

Just as one must consider how asymmetries in lower-level processes modulate asymmetries inhigher-level processes when hypothesizing about the higher-level processes, one must also consider thegoals that the lower-level processes accomplish, as well as the mappings between inputs and outputs thatthey achieve, in order to explicate the functions of the neural subsystems that underlie the lower-levelprocesses. The functions of relatively independent AVF and SVF subsystems may provide an explanationfor the observed spatial-frequency asymmetries, one that is capable of explaining why these asymmetriesevolved in the first place, but only after inputs, outputs, and their couplings are examined together.Indeed, consideration of all of the available constraints may be essential when doing the very difficult jobof stretching the brain at its interconnections.


REFERENCES

Artola, A., & Singer, W. (1987). Long-term potentiation and NMDA receptors in rat visual cortex.Nature, 330, 649-652.

Ballard, D. H. (1986). Cortical connections and parallel processing: Structure and function. Behavioraland Brain Sciences, 9, 67-120.

Banich, M. T. (1995). Interhemispheric processing: Theoretical considerations and empiricalapproaches. In R. J. Davidson & K. Hugdahl (Eds.), Brain asymmetry. Cambridge, MA: MITPress.

Banich, M. T., & Belger, A. (1990). Interhemispheric interaction: How do the hemispheres divide andconquer a task? Cortex, 26, 77-94.

Belger, A., & Banich, M. T. (1992). Interhemispheric interaction affected by computational complexity.Neuropsychologia, 30, 923-929.

Bever, T. G. (1980). Broca and Lashley were right: Cerebral dominance is an accident of growth. In D.Kaplan & N. Chomsky (Eds.), Biology and language (pp. 186-230). Cambridge, MA: MIT Press.

Biederman, I., & Cooper, E. E. (1992). Size invariance in visual object priming. Journal of ExperimentalPsychology: Human Perception and Performance, 18, 121-133.

Bowers, J. S. (in press). Different perceptual codes support priming for words and pseudowords: WasMorton right all along? Journal of Experimental Psychology: Learning, Memory, and Cognition.

Bradshaw, J. L., & Nettleton, N. C. (1981). The nature of hemispheric specialization in man. Behavioraland Brain Sciences, 4, 51-91.

Buckner, R. L., Petersen, S. E., Ojemann, J. G., Miezin, F. M., Squire, L. R., & Raichle, M. E. (1995).Functional anatomical studies of explicit and implicit memory retrieval tasks. Journal ofNeuroscience, 15, 12-29.

Burgund, E. D., & Marsolek, C. J. (1996, April). Letter-case specific priming in the right cerebral hemisphereon a modified perceptual identification task. Presented at the Cognitive Neuroscience SocietyAnnual Meeting, San Francisco, CA.

Cave, C. B., & Squire, L. R. (1992). Intact and long-lasting repetition priming in amnesia. Journal ofExperimental Psychology: Learning, Memory, and Cognition, 18, 509-520.

Christman, S., Kitterle, F. L., & Hellige, J. B. (1991). Hemispheric asymmetry in the processing ofabsolute versus relative spatial frequency. Brain and Cognition, 16, 62-73.

Churchland, P. S., & Sejnowski, T. J. (1992). The computational brain. Cambridge, MA: MIT Press.Cohen, N. J., & Eichenbaum, H. (1993). Memory, amnesia, and the hippocampal system. Cambridge, MA:

MIT Press.Cooper, L. A., Schacter, D. L., Ballesteros, S., & Moore, C. (1992). Priming and recognition of

transformed three-dimensional objects: Effects of size and reflection. Journal of ExperimentalPsychology: Learning, Memory, and Cognition, 18, 43-57.

Corballis, M. C. (1989). Laterality and human evolution. Psychological Review, 96, 492-505.Damasio, A. R., & Damasio, H. (1983). The anatomic basis of pure alexia. Neurology, 33, 1573-1583.Diamond, R., & Carey, S. (1986). Why faces are and are not special: An effect of expertise. Journal of

Experimental Psychology: General, 115, 107-117.Dimond, S. J., & Beumont, G. (1971). Use of two cerebral hemispheres to increase brain capacity.

Nature, 232, 270-271.Dimond, S. J., Gibson, A. R., & Gazzaniga, M. S. (1972). Cross field and within field integration of visual

information. Neuropsychologia, 10, 379-381.


Farah, M. J. (1990). Visual agnosia: Disorders of object recognition and what they tell us about normalvision. Cambridge: MIT Press.

Farah, M. J. (1991). Patterns of co-occurrence among the associative agnosias: Implications for visualobject representation. Cognitive Neuropsychology, 8, 1-19.

Farah, M. J. (1994). Neuropsychological inference with an interactive brain: A critique of the localityassumption. Behavioral and Brain Sciences, 17, 43-104.

Felleman, D. J., & Van Essen, D. C. (1991). Distributed hierarchical processing in primate visual cortex.Cerebral Cortex, 1, 1-47.

Fox, P. T., Miezin, F. M., Allman, J. M., Van Essen, D. C., & Raichle, M. E. (1987). Retinotopicorganization of human visual cortex mapped with positron emission tomography. Journal ofNeuroscience, 7, 913-922.

Geffen, G., Bradshaw, J. L., & Wallace, G. (1971). Interhemispheric effects on reaction time to verbal andnonverbal visual stimuli. Journal of Experimental Psychology, 85, 415-422.

Goodale, M. A., & Milner, A. D. (1992). Separate visual pathways for perception and action. Trends inNeuroscience, 15, 20-25.

Gross, C. G., Rocha-Miranda, C. E. & Bender, D. B. (1972) Visual properties of neurons ininferotemporal cortex of the macaque. Journal of Neurophysiology, 35, 96-111.

Haxby, J. V., Grady, C. L., Horwitz, B., Ungerleider, L. G., Mishkin, M., Carson, R. E., Herscovitch, P.,Schapiro, M. B., & Rapoport, S. I. (1991). Dissociation of object and spatial visual processingpathways in human extrastriate cortex. Proceedings of the National Academy of Sciences USA, 88,1621-1625.

Hellige, J. B. (1983). Hemisphere x task interaction and the study of laterality. In J. B. Hellige (Ed.),Cerebral hemispheric asymmetry: Method, theory, and research (pp. 411-443). New York: Praeger

Hellige, J. B. (1993a). Hemispheric asymmetry: What's right and what's left. Cambridge, MA: HarvardUniversity Press.

Hellige, J. B. (1993b). Unity of thought and action: Varieties of interaction between the left and rightcerebral hemispheres. Current Directions in Psychological Science, 2, 21-25.

Hinton, G. E., McClelland, J. L., and Rumelhart, D. E. (1986). Distributed representations. In D. E.Rumelhart & J. L. McClelland (Eds.), Parallel distributed processing: Explorations in themicrostructure of cognition. Volume 1: Foundations (pp. 77-109). Cambridge, MA: MIT Press.

Hummel, J. E., & Biederman, I. (1992). Dynamic binding in a neural network for shape recognition.Psychological Review, 99, 480-517.

Jacobs, R. A., & Kosslyn, S. M. (1994). Encoding shape and spatial relations: The role of receptive fieldsize in coordinating complementary representations. Cognitive Science, 18, 361-386.

Kitterle, F. L., Christman, S., and Hellige, J. B. (1990). Hemispheric differences are found in theidentification, but not the detection, of low versus high spatial frequencies. Perception &Psychophysics, 48, 297–306.

Kitterle, F. L., & Selig, L. M. (1991). Visual field effects in the discrimination of sine-wave gratings.Perception & Psychophysics, 50, 15-18.

Knapp, A., & Anderson, J. A. (1984). A signal averaging model for concept formation. Journal ofExperimental Psychology: Learning, Memory, and Cognition, 10, 616-637.

Koivisto, M. (1995). On functional brain asymmetries in perceptual priming. Brain and Cognition, 29,36-53.

Komatsu, Y., Fujii, K., Maeda, J., Sakaguchi, H., & Toyama, K. (1988). Long-term potentiation of synaptictransmission in kitten visual cortex. Journal of Neurophysiology, 59, 124-141.


Kosslyn, S. M. (1987). Seeing and imagining in the cerebral hemispheres: A computational approach.Psychological Review, 94, 148-175.

Kosslyn, S. M., Chabris, C. F., Marsolek, C. J., & Koenig, O. (1992). Categorical versus coordinate spatialrelations: Computational analyses and computer simulations. Journal of ExperimentalPsychology: Human Perception and Performance, 18, 562-577.

Kosslyn, S. M., & Koenig, O. (1992). Wet mind: The new cognitive neuroscience. New York: Free Press.Kosslyn, S. M., Alpert, N. M., Thompson, W. L., Maljkovic, V., Weise, S. B., Chabris, C. F., Hamilton, S. E.,

Rauch, S. L., Buonanno, F. S. (1993). Visual mental imagery activates topographically organizedvisual cortex: PET investigations. Journal of Cognitive Neuroscience, 5, 263-287.

LaBerge, D. (1995). Computational and anatomical models of selective attention in object identification.In M. S. Gazzaniga (Ed.), The cognitive neurosciences (pp. 649-663). Cambridge, MA: MIT Press.

Lehky, S. R. & Sejnowski, T. J. (1988). Network model of shape-from-shading: Neural function arisesfrom both receptive and projective fields. Nature, 333, 452-454.

Levine, D., & Calvanio, R. (1989). Prosopagnosia: A defect in visual configural processing. Brain andCognition, 10, 149-170.

Marr, D. (1982). Vision. San Francisco, CA: W. H. Freeman.Marshall, J. C. (1981). Hemispheric specialization: What, how, and why. Behavioral and Brain Sciences,

4, 72-73.Marsolek, C. J. (1994, March). Computational incompatibility for processing specific- vs. abstract-visual-

form information. Presented at the Cognitive Neuroscience Society Annual Meeting, SanFrancisco, CA.

Marsolek, C. J. (1995). Abstract-visual-form representations in the left cerebral hemisphere. Journal ofExperimental Psychology: Human Perception and Performance, 21, 375-386.

Marsolek, C. J., Ecker, B. C, Ewald, B. J., Stoltz, E. L., Wright, M. E., & Nicholas, C. D. (1995, November).Letter-case similarity and backward pattern masking modulate form-specific priming in the cerebralhemispheres. Presented at the Psychonomic Society Annual Meeting, Los Angeles, CA.

Marsolek, C. J., Kosslyn, S. M., & Squire L. R. (1992). Form-specific visual priming in the right cerebralhemisphere. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 492-508.

Marsolek, C. J., Schacter, D. L., & Nicholas, C. D. (1996). Form-specific visual priming for newassociations in the right cerebral hemisphere. Memory & Cognition, 24, 539-556.

Marsolek, C. J., Squire, L. R., Kosslyn, S. M., & Lulenski, M. E. (1994). Form-specific explicit and implicitmemory in the right cerebral hemisphere. Neuropsychology, 8, 588-597.

Mazzoni, P., Andersen, R. A., & Jordan, M. I. (1991). A more biologically plausible learning rule forneural networks. Proceedings of the National Academy of Sciences USA, 88, 4433-4437.

McClelland, J. L., McNaughton, B. L., & O'Reilly, R. C. (1995). Why there are complementary learningsystems in the hippocampus and neocortex: Insights from the successes and failures ofconnectionist models of learning and memory. Psychological Review, 102, 419-457.

McClelland, J. L., & Rumelhart, D. E. (1985). Distributed memory and the representation of general andspecific information. Journal of Experimental Psychology: General, 114, 159-188.

Nicholas, C. D., & Marsolek, C. J. (1996, April). Differential interhemispheric transfer for abstract andspecific visual-form information. Presented at the Cognitive Neuroscience Society AnnualMeeting, San Francisco, CA.

O'Reilly, R. C., Kosslyn, S. M., Marsolek, C. J. & Chabris, C. F. (1990). Receptive field characteristics thatallow parietal lobe neurons to encode spatial properties of visual input: A computationalanalysis. Journal of Cognitive Neuroscience, 2, 141-155.


Petersen, S. E., & Fiez, J. A. (1993). The processing of single words studied with positron emissiontomography. Annual Review of Neuroscience, 16, 509-530.

Plaut, D. C., & Hinton, G.E. (1987). Learning sets of filters using back-propagation. Computer Speechand Language, 2, 35-61.

Rhodes, G. (1985). Lateralized processes in face recognition. British Journal of Psychology, 76, 249-271.Robertson, L. C., & Lamb, M. R. (1991). Neuropsychological contributions to theories of part/whole

organization. Cognitive Psychology, 23, 299-330.Rueckl, J. G., Cave, K. R., & Kosslyn, S. M. (1989). Why are “what” and “where” processed by separate

cortical systems? A computational investigation. Journal of Cognitive Neuroscience, 1, 171-186.Rumelhart, D. E., Hinton, G., & Williams, R. (1986). Learning internal representations by error

propagation. In D. E. Rumelhart & J. L. McClelland (Eds.), Parallel distributed processing:Explorations in the microstructure of cognition. Volume 1: Foundations (pp. 318-362).Cambridge, MA: MIT Press.

Schacter, D. L., Reiman, E., Uecker, A., Polster, M. R., Yun, L. S., & Cooper, L. A. (1995). Brain regionsassociated with retrieval of structurally coherent visual information. Nature, 376, 587-590.

Schacter, D. L., & Tulving, E. (Eds.). (1994). Memory systems 1994. Cambridge, MA: MIT Press.Schwartz, E. L., Desimone, R., Albright, T. D., & Gross, C. G. (1983). Shape recognition and inferior

temporal neurons. Proceedings of the National Academy of Sciences USA, 80, 5776-5778.Sergent, J. (1982). The cerebral balance of power: Confrontation or cooperation? Journal of

Experimental Psychology: Human Perception and Performance, 8, 253–272.Sergent, J. (1989). Structural processing of faces. In A. W. Young & H. D. Ellis (Eds.), Handbook of

research on face processing (pp. 57-91). Amsterdam: Elsevier.Sergent, J., & Bindra, D. (1981). Differential hemispheric processing of faces: Methodological

considerations and reinterpretation. Psychological Bulletin, 89, 541-554.Sergent, J., Ohta, S., & MacDonald, B. (1992). Functional neuroanatomy of face and object processing.

Brain, 115, 15-36.Squire, L. R. (1992). Memory and the hippocampus: A synthesis from findings with rats, monkeys, and

humans. Psychological Review, 99, 195-231.Squire, L. R., Ojemann, J. G., Miezin, F. M., Petersen, S. E., Videen, T. O., & Raichle, M. E. (1992).

Activation of the hippocampus in humans: A functional anatomical study of memory.Proceedings of the National Academy of Sciences USA, 89, 1837-1841.

Tarr, M. J. (1995). Rotating objects to recognize them: A case study on the role of viewpointdependency in the recognition of three-dimensional objects. Psychonomic Bulletin & Review, 2,55-82.

Tootell, R. B. H., Silverman, M. S., Switkes, E., & De Valois, R. L. (1982). Deoxyglucose analysis ofretinotopic organization in primate striate cortex. Science, 218, 902-904.

Ungerleider, L. G. & Mishkin, M. (1982). Two cortical visual systems. In D. J. Ingle, M. A. Goodale, & R.J. W. Mansfield (Eds.), The analysis of visual behavior (pp. 549-586). Cambridge, MA: MIT Press.

Van Kleeck, M. H. (1989). Hemispheric differences in global versus local processing of hierarchical visualstimuli by normal subjects: New data and a meta-analysis of previous studies. Neuropsychologia,27, 1165-1178.

Zipser, D. & Andersen, R. A. (1988). A back-propagation programmed network that simulates responseproperties of a subset of posterior parietal neurons. Nature, 331, 679–684.


AUTHOR NOTES

Preparation of this chapter was supported by the National Institute of Mental Health, GrantMH53959-01; by the McDonnell-Pew Cognitive Neuroscience Center and the Arizona Cognitive ScienceProgram of the University of Arizona; and by the Center for Research in Learning, Perception, &Cognition in conjunction with the National Science Foundation (GER 9454163), the Office of the VicePresident for Research, and Dean of the Graduate School of the University of Minnesota.

Correspondence may be sent to C. J. Marsolek, Department of Psychology, University ofMinnesota, 75 East River Road, Minneapolis, MN, 55455. Electronic mail may be sent to:[email protected].


FIGURE CAPTIONS

Figure 1. Different specific instances that belong to the same abstract category of visual form.

Figure 2. In the upper display, the relatively invariant information (right) that is common to differentspecific instances in the same abstract category of visual form (left). In the lower display, the visuallydistinctive information (right) that distinguishes specific instances in the same abstract category of visualform (left).

Figure 3. AVF mappings of input forms (left) to output representations (right) simulated in neuralnetwork models.

Figure 4. SVF mappings of input forms (left) to output representations (right) simulated in neuralnetwork models.

Figure 5. Intermediate-level mappings of input forms (left) to output representations (right) simulated inneural network models.


BEARBEARBEARBEARbearbearbearbear



Relatively Invariant Information



Visually Distinctive Information


Output 1

BEARBEARBEARBEAR

bearbearbearbear

GANGGANGGANGGANG

ganggangganggang

AVF Mappings

Output 2


BEARBEARBEARBEAR

bearbearbearbear

GANGGANGGANGGANG

ganggangganggang

SVF Mappings

Output 1Output 2Output 3Output 4





Output 1BEARBEARBEARBEAR

bearbearbearbear

GANGGANGGANGGANG

ganggangganggang

Intermediate-Level Mappings

Output 4

Output 2

Output 3

in visual-form recognitionlevels.psych.umn.edu/pubs_index/pubs_web_pdfs/m_b_97.pdf ·...

Documents